Comparative Analysis of LSTM, BiLSTM, Stacked LSTM, and Attention-LSTM for Multi-Resolution Rainfall Forecasting in Bali

Abstract

Accurate rainfall forecasting remains challenging in tropical islands such as Bali, Indonesia, where localized convective rainfall limits the reliability of physics-based models. While deep learning architectures based on Long Short-Term Memory (LSTM) are increasingly used for rainfall forecasting, most prior studies compare LSTM variants at a single temporal resolution without a classical baseline, leaving it untested whether their reported advantage holds across resolutions or is an artifact of the resolution chosen. This study evaluates four recurrent architectures — LSTM, BiLSTM, Stacked LSTM, and Attention-LSTM — for univariate rainfall forecasting at daily, decadal, and monthly resolutions, using 26 years (2000–2025) of observations from a BMKG station in Bali, benchmarked against three classical baselines (Persistence, Climatology, and ARIMA) and tuned via random search. Results show that inter-architecture differences within a single resolution are small and inconsistent, with no architecture winning across all resolutions: Stacked LSTM is marginally best at daily (RMSE 15.26 mm/day) and decadal (73.53 mm/decade) scales, while BiLSTM leads at the monthly scale (154.64 mm/month). More critically, the deep learning models' advantage over classical baselines is strongly resolution-dependent: substantial at daily and decadal resolutions, but nearly indistinguishable from Persistence and ARIMA at the monthly resolution, where only 312 training sequences are available. These findings indicate that model selection for rainfall forecasting should be conditioned on temporal resolution and data availability rather than treated as a fixed architectural choice, and that simpler LSTM models offer a more computationally efficient default than their more complex variants for tropical, data-limited settings.